Claude Code 6 Seviyede Öğren: Kodlamadan Yazılım Yap

Technological Concepts & What the Video Teaches (Cloud Coding → “Level 6+”)

Cloud Code as an AI coding agent

Cloud Code is presented as an AI coding agent that operates your project in the terminal. Unlike chat-only tools (which mostly take input and produce output), it can:

• Read your project files
• Write code
• Run, correct, and rewrite code
• Deploy changes using Git for testing and deployment

Designed for non-programmers

The narrator emphasizes that you don’t need to code. If you can clearly describe what you want, Cloud Code can generate:

• Working websites
• Prototypes
• Fixes

Multi-level workflow model (“Commander → Planner → Context Engineer → Tools/Skills → Orchestrator”)

The course is structured to upgrade how you use the system—from basic prompting to advanced agent orchestration.

---

Key Product Capabilities / Features Highlighted

• Website/app generation from scratch
  • Landing pages, dashboards, e-commerce layouts
• Prototype generation
  • For mobile/web app ideas
• Project debugging and fixing
  • Inspect a project, find issues, test, and release fixes
• Integrations inside projects
  • Mentions API integrations
• Automated test and documentation generation
• Uploads and edits via Git repositories
  • GitHub/GitLab-style workflow
• Model switching during work
  • Use faster/economical models vs stronger models when needed

---

Installation / Setup Guides (Step-by-Step)

The video explains multiple ways to set up Cloud Code:

1) Terminal install method

• Run a single-line command (macOS/Windows/Linux)
• Notes Node.js usage as needed for an alternative install method (npm)
• Suggests installing Node.js via Google (as mentioned in the video)

2) VS Code extension

• Install Cloud Code as a VS Code extension
• The extension also supports terminal usage without a separate terminal setup step

3) Desktop app

• Download and use a local desktop interface (chat/chat+code/context style)
• The narrator says the course won’t rely on the desktop app in that video, but it exists

---

Account + Pricing / Limits (Analysis Included)

• Cloud Code requires a paid plan or API credits
• A free plan doesn’t run Cloud Code
• Plans mentioned include:
  • Pro
  • Max
  • Team
  • Enterprise
  • API credits

Token pricing logic (explained)

• Input tokens vs output tokens are priced per million (example numbers are given in the video)

Advice included

• API pay-as-you-go can become more expensive than Max for monthly usage due to daily/continuous billing
• Max plan includes a high token quota (quota resets when reached)

---

Models and When to Use Them

Models mentioned

• Haiku (fast/economical)
• Sonnet (strong default)
• Opus (deep/architectural) (Naming may be slightly off due to subtitles.)

Strategy

• Use stronger models (e.g., Sonnet/Opus) for harder parts
• Switch models dynamically (e.g., via slash-model commands)

---

Cloud Code Commands (Tutorial-like Command Set)

Core commands appear repeatedly across the “leveling” approach:

Planning and session management

• Plan mode: plan (step-by-step planning before coding)
• Clear: wipes conversation history for a fresh start
• Compact: summarizes current context so you can continue later without copying everything verbatim

Context/cost inspection

• Commands such as:
  • context window
  • slashcost
  • usage
• Used to track tokens and cost

Model switching

• Example: slash model
• Or slash commands like slash opus/haiku/sonnet

Project initialization

• slash init creates a project-specific Cloud MD file

Other capabilities mentioned

• Remote control via QR
• Permissions/approval workflow for terminal commands
• Batch/agent-related commands
• “Interrupt/questions mid-process”

---

The “Levels” (What Changes from Level to Level)

Level 0: Foundation — Understand Cloud Code + Setup + Basic Workflow

• How Cloud Code differs from standard chat:
  • It operates in the terminal
  • It reads files
  • It uses Git for deployment/testing

Level 1: Commander — Prompting without Context

• Mistake described:
  • Using the most powerful model immediately and asking something like “create a landing page” without context
• Result:
  • Generic UI/sections that look “classic AI-made” (e.g., stereotypical button styles) because the model guesses
• Skills to practice:
  • Write clear, specific instructions (audience, purpose, CTAs, contact info, goals)
  • Review outputs (front-end structure and terminal logs)
  • Basic “terminal literacy” to understand what Cloud is doing

Level 2: Planner — Use Plan Mode + Structured Pre-Design

• Correct approach: plan before implementation
• The planner can ask for:
  • Language
  • Structure (single vs multi-page)
  • Theme
  • Technology choices
• Practical artifact:
  • Reference files (e.g., using an @ reference to a company profile text file)
• Trap warned:
  • Even with planning, Cloud can still speculate if context is insufficient—so you must guide and challenge it

Level 2.5 (Implied): “Devil’s Advocate” Questioning

The narrator suggests you should actively challenge the agent by asking things like:

• Do the proposed decisions actually make sense?
• Can the agent critique alternatives instead of blindly executing?

Level 3: Context Engineering

• Introduces Cloud MD file
  • A structured “instruction memory” containing:
    • Project description, goals
    • Relevant tech, standards, keywords
• Key thesis:
  • You act as a curator/director, and Cloud Code becomes the implementer
• Major caution: don’t bloat context
  • More context files can worsen performance (research/agent studies are referenced)
  • Risks include higher token costs and hallucinations
• Context management commands:
  • Clear vs Compact
  • Notes about context window limits and contextual decay
• Best practice recommended:
  • Summarize key info → copy to a local file → clear → paste/use that summary later

Level 4: Connection Engineering (Tooling / MCP)

• Adds external tools via MCP (Model Context Protocol)-style connections
• Example connected systems mentioned:
  • Supabase (database)
  • Figma (design templates/assets)
  • Playwright (browser automation, implied by naming)
• Authentication/API keys:
  • The cloud must know whose account to connect to; tokens/keys may be required
• Trap warned:
  • “Candy store trap”: too many tools increases the chance the AI chooses the wrong one and makes worse decisions
• Introduces frameworks:
  • “Blueprint” code/workflows from frameworks; may or may not be necessary

Level 5: Skills (Custom and Curated Tool Use)

• “Skills” are reusable agent instructions stored as files in the cloud (MD-style)
• Avoid overloading:
  • Too many MCPs/skills can worsen decisions and create token/context overhead
• Skills management:
  • Select only relevant skill modules (e.g., Supabase workflow skills)
  • Mentions creating your own skills via a skill creator plugin
• Skill composer concept:
  • Combine multiple skills (A, B, C) into a new consolidated skill

Level 6: Orchestrator / Agent Teams (Parallel Work + Coordinated Communication)

The video argues that multiple agents need coordination, not just parallel sessions.

Three structures discussed

1. Multiple terminals / separate agents
   • Parallel, but chaotic if roles aren’t managed
2. Work trees / sub-agents
   • Cloud creates intermediate work processes; sub-agents feed outputs back
3. Agent Team (coordinated agents)
   • A team leader assigns tasks; agents communicate through the system

Git repository “work tree” and role separation

• Use Git to structure outputs so agents don’t overwrite each other

Trade-off

• Agent Teams cost more tokens due to internal communication
• Recommend monitoring cost/context during agent-team runs

Enabling Agent Teams (tutorial-style)

• Use natural-language prompts such as:
  • “enable agent team”
  • “create 3 agents”
• Agents might be split by roles such as:
  • frontend, backend, authentication, testing/writer
• Settings file updates happen automatically (as described)

---

Overall Review / Analysis Takeaways Emphasized in the Video

• Cloud Code success depends less on the “most powerful model” and more on:
  • Context quality
  • Planning before coding
  • Controlled context window management
  • Tool/skill selection precision
  • Agent orchestration with division of labor
• Major failure mode:
  • Skipping levels → teams that generate “AI slop” because fundamentals weren’t learned
• Competitive/market analysis claim:
  • People who learn this early will be advantaged
  • The narrator claims familiarity is rare globally and suggests Turkey is even behind

---

Main Speakers or Sources

• Main speaker
  • A single narrator/instructor throughout the video (no other consistent named host)
• External referenced sources (mentioned as docs/studies, not direct speakers)
  • “Cloud Code official documentation” (for Agent Team usage and Cloud MD guidance)
  • Research references discussed for context bloat/performance decay (names like “Kroma” and an AI-agent study are mentioned, but not quoted as primary speakers)